AI research organization “Human-Centered AI Institute”

This time, I would like to provide you with some insights based on my various experiences, such as supporting the industry at a consulting firm, leading research and development at an internet company, and leading the DX and AI fields at a global professional firm. The image below shows the Human-Centered AI Institute, an AI research organization launched by the Hakuhodo DY Group in April 2024.

AI has shown remarkable technological advances in recent years, and while it has an impact on all fields and industries, it has also sparked discussions about risks and ethical issues. Therefore, the concept that "AI is intended to be useful to humans and should be used in that way" is beginning to be given great importance. This is the key point of this presentation, "Human-Centered AI."

The basis of Human-Centered AI is the idea that AI should be developed to benefit society, for example by satisfying human needs or by appropriately reflecting human perspectives in the AI development process. We believe that we should take the path beyond that, that is, the relationship between AI and creativity, to the next step. To do this, we need to create AI that changes people and society, while thoroughly exploring how AI can unleash the various potential and creativity that humans possess. The organization that will make this happen is the Human-Centered AI Institute.

What is Human-Centered AI?

The following diagram shows the concept of Human-Centered AI management and updates.

This diagram has several layers. At the very bottom are so-called AI ethics, that is, principles for developing and using AI. These are supported by elements such as accountability, transparency, fairness, and safety, which are being discussed in various places.

Above that is a layer for process and operation management. Items include components for making AI interpretable and explainable, reflecting human values through human-participation development, and data governance for the preparation and management of the underlying data. Above that are components for overseeing these and applying AI governance to create responsible AI.

We believe that we should create a layer of creativity and society on top. Here, we position human-centered design components that make things easy for people to use, and place the area of human-AI interaction, which has become a hot topic of research in recent years.

It is also important to ensure engagement with stakeholders related to AI. AI developers and services need to manage their development and utilization by involving stakeholders, including not only AI users but also consumers and society at large. By reflecting this, Human-Centered AI should be created.

The Hakuhodo DY Group is putting Human-Centered AI into practice through two structures. One of these is the Human-Centered AI Institute, which I introduced at the beginning of this article, where cutting-edge research and development organizations are creating future roadmaps and working on alliances with various companies.

The other is the Human-Centered AI Initiative, a group-wide project. AI leaders from each Hakuhodo DY Group company are also participating in this initiative, and after sharing information on best practices and AI regulations, the group is discussing how to work together to build governance. Furthermore, the group aims to use AI to improve system efficiency and productivity, and is exploring how to provide new value to clients and consumers rather than simply using AI itself.

AI Technology Trends

Various information about AI is reported in the news every day, but it is difficult to see what the future holds for AI. In this situation, the following diagram summarizes the contents of our discussion.

First, the vertical axis shows the degree to which research is academic (specialized) or practical. Modern AI is used in practice, and the know-how discovered is fed back into the research field, which can lead to technological advances. Here, this is defined as practical. Next, the horizontal axis shows the degree to which the research is process-oriented, which aims to streamline the process, and experience-oriented, which emphasizes how the users and those who receive the services can have a new experience.

Looking at the diagram, we can see that while there is a major trend from process-oriented to experience-oriented as in ① to ⑤, the technology itself shifted from academic to practical, and then returned to academic. From here on, we will introduce the characteristics of ① to ⑤ in order.

Generative AI

In the case of generative AI, text-based generative AI will become a big trend in 2023, and multimodal AI such as images and videos will develop. Currently, multimodal long context is attracting attention, and for example, NDIVIA, a semiconductor player that can be said to be the driving force behind generative AI technology, offers high-quality generative AI. In addition, according to a research report, more than 70% of listed companies in Japan have introduced generative AI as a foundation and are linking it with their own data.

When I asked information system personnel, digital transformation personnel, CDOs, and CDXOs from over 100 companies about how they use generative AI, I found that there is a wide range of use cases, regardless of whether it is in the back office or the front office, as shown in the following diagram. As of 2023, there was no company that covered all of the elements in the diagram, but the use of generative AI has progressed to the point that it would not be surprising if such a company existed today.

Specific uses include financial data analysis and improving the accuracy of stock price predictions. Regarding the latter, a paper published in April 2023 described how "news headlines can be extracted and passed to ChatGPT, and an answer can be obtained by asking whether the stock price will rise or fall as a result," and "various news items can be input, and based on the answers obtained from ChatGPT, the prediction model can be revised using previous stock price data."

Given this background, the question now is how your company is using generative AI and what use cases it is developing.

Enterprise AI

② Enterprise AI focuses on how generative AI is used in conjunction with in-house systems. Most companies are probably working on this with a vision like the one below.

Normally, employees and various systems and databases are separate. However, by creating a generative AI portal and making inquiries and instructions in natural language, it becomes possible to obtain answers linked to various systems, databases, and documents, and to process business operations. As a result, someone who was previously only responsible for accounting work may be able to take on general affairs, human resources, or even more. This mechanism in which generative AI becomes a hub is called RAG (Retrieval Augmented Generation).

Hakuhodo DY Group is building a "multi-agent brainstorming AI" that applies the RAG system. When planning a product, there is a process of collaborating with various ideas before delivering it to the market, and with the power of RAG, this process can be improved in quality while controlling risks and speeding up the process.

A distinctive feature of this system is the introduction of personas using generative AI. For example, experts in manufacturing, logistics, and public relations can be created based on generative AI, and these can be used to conduct product development and reviews and discussions for the GoTo Market. In fact, some of our clients have linked their researcher's database, know-how, and management systems, and by introducing the researcher's persona into the system and holding brainstorming sessions, they have been able to realize quality product development discussions.

And when using generative AI to increase productivity and efficiency in this way, the key to success is the "human-centered AI" we mentioned at the beginning. Traditionally, people have taken the approach shown on the left side of the diagram below, of "replacing the work process performed by employees with AI from this point to this point."

However, modern AI, especially generative AI, is basically based on probability and statistical processing based on machine learning. Therefore, there is always an error in the answer. In addition, in generative AI, the output is creative, so the deviation is even greater, and in some contexts it is called hallucination (false information). In other words, the approach of taking business processes from people and automating them does not assume that there will be errors or deviations as a premise, so there are problems that do not always work well.

It's a mistake to expect 100% accurate automation from AI, and we should actually move towards human-centered AI, as shown on the right side of the diagram. In some cases, generative AI will come up with various ideas, and humans will use them to come up with further ideas, while in other cases, AI will act as a foundation to generate collaboration between workers. In this way, utilizing generative AI in a way that unleashes creativity is important from an AI technology perspective, and I think it is important for companies to take an approach based on this.

However, the winds of complexity are a problem that stands in the way of achieving this. When aiming to improve productivity using RAG, you can get reasonably good results at first. However, when you expand the area or scope from there, you are pushed back by the complexity of the system, and it doesn't work out well. In a situation where trying to expand the area reduces overall performance, how to overcome this wind of complexity is a very important point.

As a result, many companies have come to the conclusion that "we need to properly consolidate our know-how on how to utilize AI," and that "if we don't properly organize our data, we can't use generative AI." As the top of the image says, "Organizational efforts by the CoE (Center of Excellence)," the accumulated AI know-how must be effectively utilized across the entire company. Recently, more and more companies are working on CoE and data management.

In particular, the importance of data management, or the so-called DMBOK (Data Management Body of Knowledge), has been strongly questioned in the past 5 to 6 years. Currently, many companies understand that while generative AI can increase people's creativity and productivity, the accuracy of such work will not improve unless data is prepared, and the promotion of data management based on the use of generative AI is becoming a trend. In addition, some companies are creating their own LLMs (large-scale language models) using their own data.

Human-AI Interaction

For now, let's put point ③ aside for now and look at point ④, Human-AI Interaction. This is the idea that "developing various interactions between humans and AI will lead to breakthroughs in AI technology." Representative examples include AR, VR, and XR, as well as AI x Metaverse and even AI x Neuroscience.

In this field, there is much discussion about how AI can change the way services are provided to end users and customers. Hakuhodo DY Group is creating a very interesting Human-AI Interaction service called "AI Rap Business Cards." It's a very interesting system in which you enter your profile information and upload a photo, and the AI automatically generates and sings a rap that highlights you. It's a kind of self-expression, so to speak.

We are also creating a service that allows you to communicate with virtual sei-katsu-sha, built through surveys and data analysis based on sei-katsu-sha ideas. Using this, we can recreate 7,000 different types of sei-katsu-sha and conduct interviews, and by watching discussions between virtual sei-katsu-sha, it is possible to gain a deeper understanding of sei-katsu-sha. We believe that this service will become a new cornerstone in terms of "how to create value by combining Hakuhodo DY Group's understanding of sei-katsu-sha with generative AI" and "what kind of value to provide to client companies and sei-katsu-sha."

This is targeted at B2C, but the development of generative AI technology has been remarkable, and various things are emerging that change the end-user touchpoint in the B2B domain as well. For example, generative AI is being used to support complex and advanced maintenance work in manufacturing factories.

When equipment breaks down on-site, it is necessary to go and fix it, but currently there is too much information, such as maintenance history and how to deal with errors, for even the workers to keep track of it all. For this reason, on-site contact is made by phone, and support staff handles the issue. However, by introducing generative AI into this, it becomes possible to handle the issue with multi-modal automated voice. Moreover, daily work reports and maintenance data are created from the content of the interactions recorded by generative AI, making it data-driven, and these are reflected in the digital twin. In this way, new forms of Human-AI Interaction are emerging that are changing the interfaces of workers.

New World

The discussion of unexplored territory is presented in ⑤ New World. Among these, one of the most important technologies is World Models. This emerged from robotics research and is used to efficiently learn the results of actions, such as turning the steering wheel to the right for a car or picking up a robot. Unlike conventional methods, a major feature of this method is that it allows a robot to act based on imagination, rather than simply waiting for the results of an action.

The world model is attracting a lot of attention as a technology that will bring about a breakthrough in AI, which is already excellent, and we believe that it is extremely important not only in the field of robotics, but also as a factor that supports highly specialized professionals. This is because professionals act by thinking several steps ahead. Even if they input a prompt into a generative AI and the result is returned, they will not be satisfied if it is only one step ahead.

In that case, we will need an approach that takes into account AI responses such as "This is XX, so what follows is XX." However, we are not yet at a level where this can be realized. This world model holds the key, and I think that discussions in this area will become very lively in the future.

Trustworthy AI

The last one is "Approach to Trustworthy AI", and this time we will introduce a probabilistic programming technique called "Probabilistic Programming".

Today's AI and LLM are based on deep learning, and are essentially black Box inside. This is not desirable when making important decisions in fields related to life and safety, so some people are of the opinion that "it would be good to be able to verify the reasoning behind the decision." The solution to this problem is Probabilistic Programming, which allows you to build AI by clarifying "what it is doing" based on Bayesian estimation, etc.

Also, while past events are digitized and trained by AI, human predictions and wisdom about future events are not included in the data. Probabilistic Programming is useful for reflecting such flexible human wisdom. It is an important technology for building trustworthy AI, including collaboration with LLM, and various research projects are being conducted.

With the recent spread of AI, various discussions on AI regulations are underway in the international community. The EU AI Act, enacted by the EU in May 2024, is known as the world's first comprehensive AI regulation. This rule classifies risks by level and takes an approach that requires strict management and transparency of AI, but it is actually applied outside the EU as well, and it is expected to have a global impact due to very severe penalties.

AI performance is steadily improving, and as a result, discussions on new regulations are underway in the United States, Japan, and other countries. How to reflect these changes is an important point in ensuring reliable AI, and I believe that the concept of Human-Centered AI that I mentioned at the beginning will play an important role in this regard.

in conclusion

I have had discussions with various experts about how the AI trends introduced here will affect consumers and society, and how they will shape the future. Based on these discussions, I have created a white paper that I am now providing to you. If you are interested, please download it from the URL below.

https://www.hakuhodody-holdings.co.jp/news/corporate/2024/07/4890.html

Hakuhodo DY Holdings Inc.
Chief AI Executive Officer Human-Centered AI Institute
Masaya Mori
After working for a consulting firm and an internet company, he joined a professional firm and is engaged in supporting companies in the digital transformation field. He is a professor at Tohoku University, an advisor for the Collaborative Platform Development at the University of Tokyo, and an advisor for the Japan Deep Learning Association.